Learning content‐adaptive feature pooling for facial depression recognition in videos

Recently, a deep representation of facial depression built on convolutional neural networks has shown impressive performance in video‐based depression recognition. However, most existing approaches either fix the weights or using a certain heuristics to integrate the frame‐level facial features, resulting in suboptimal feature aggregation in encoding the helpful while discarding noisy information in videos. To address this issue, the authors introduce the memory attention mechanism in a regression network to learn a deep discriminative depression representation, where the residual network module aims at learning frame‐level deep feature, while the attention module acts as a pooling layer by adaptively learning the weights emphasising or suppressing face images with varying poses and imaging conditions. They empirically evaluate the proposed approach on a benchmark depression dataset, and the results demonstrate the superiority of their approach over the state‐of‐the‐art alternatives.